Is a non-generative, retrieval-only architecture (excluding LLMs from the search steps) the optimal solution for building highly reliable and cost-effective personal memory systems?